A method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks

Abstract

This invention discloses a method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks (GANs). The method first fixes the network parameters of the generator, using training sample data as real samples and scattering data generated by the generator as fake samples. The real and fake samples are combined according to a predetermined ratio and then input into a discriminator and classifier for iterative training until the training termination condition is met. Then, the network parameters of the discriminator network are fixed, and the scattering data generated by the generator is used as real samples, again input into the discriminator and classifier for iterative training until the training termination condition is met. The trained GAN then predicts the scattering data of vortex electromagnetic waves on targets with corresponding parameters based on the input target parameters and random noise. This invention fills the gap in the prediction of vortex electromagnetic wave target scattering characteristics, improves the efficiency of data prediction, and reduces memory overhead.

Description

Technical Field

[0001] This invention relates to the field of radar target characteristic simulation, and in particular to a method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks. Background Technology

[0002] Determining the electromagnetic scattering distribution of a target based on its known geometric and physical parameters is an important research direction in electromagnetic scattering. In recent years, with the development of deep learning technology, many scholars have incorporated it into solving this problem. Currently, most research in this area is based on plane waves. However, vortex electromagnetic waves, due to their unique phase distribution and amplitude characteristics, exhibit different scattering characteristics compared to plane waves. Unlike traditional plane waves that illuminate targets from multiple consecutive angles, vortex electromagnetic waves create differentially distributed electromagnetic excitations at different targets within the same beam distance. This is expected to allow the target echo to contain more target information. To widely apply vortex electromagnetic waves to radar detection and identification, the speed of acquiring vortex electromagnetic wave scattering data needs to be addressed. Therefore, how to obtain vortex electromagnetic wave scattering data more quickly, and how to predict vortex wave electromagnetic scattering characteristics and generate scattering data based on deep learning, have become important research directions.

[0003] In the inversion of electromagnetic scattering characteristic parameters of plane waves, Li Chao et al. proposed a rapid modeling method, which studied the electromagnetic scattering problems of volumes and surfaces, and media and metals. This method utilizes multiple decomposition techniques of iterative algorithms and special basis function methods based on direct solution to achieve rapid solution of scattering characteristics. The paper also investigated the electromagnetic inverse scattering problem of targets placed on rough surfaces and used machine learning methods to invert the geometric characteristic parameters of targets on rough surfaces. However, the paper did not address the prediction of the scattering characteristics of targets by vortex electromagnetic waves in the microwave band. Since the electromagnetic excitation distributions of plane waves and vortex waves on targets are different, the method studied cannot be applied to the prediction of the scattering characteristics of targets by vortex electromagnetic waves.

[0004] In addition, Li Shiyi et al. proposed an augmentation method based on a pyramid-shaped multi-scale generative adversarial network in their paper "SAR Image Sample Augmentation Based on Multi-Scale Generative Adversarial Networks." This method employs a pyramid structure to train the same image at different sizes. By using the Inception (heuristic network) module as the front end of the generator, it effectively extracts features from the image at different sizes. Furthermore, this method utilizes residual-dense modules to fully leverage image features, thereby improving the accuracy of the augmented images. Adding these augmented images to the dataset to train an object detection model yields a better-performing model. However, this method can only train on a single image at a time, and retraining is required when augmenting new images, leading to high training time and cost.

[0005] In general, due to their unique phase characteristics and amplitude distribution, vortex electromagnetic waves exhibit significantly different electromagnetic scattering characteristics compared to plane waves. The polarization, amplitude, and phase of vortex electromagnetic waves incident on different surface elements of a target vary, rendering traditional plane wave target scattering prediction methods inapplicable. Current prediction techniques for target electromagnetic scattering characteristics are only applicable to plane wave incident conditions; techniques for predicting the target scattering characteristics of vortex electromagnetic waves require further research. Furthermore, current intelligent networks that utilize target electromagnetic scattering characteristics for prediction and identification primarily rely on traditional high-frequency modeling and simulation, a process that typically involves complex numerical simulations and iterative corrections, consuming substantial time and computational resources. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks, thereby improving the efficiency of predicting vortex electromagnetic wave scattering characteristic data while reducing resource consumption, in order to address all or part of the problems mentioned above.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks, comprising:

[0009] The target parameters and random noise are input into a trained generative adversarial network to generate vortex electromagnetic wave scattering data on the target; among which,

[0010] The generative adversarial network includes a generator, a discriminator, and a classifier;

[0011] The generator is configured to generate vortex electromagnetic wave scattering data of the target based on the input target parameters and random noise.

[0012] The discriminator is configured to: distinguish between genuine and fake input training sample data or scattering data generated by the generator;

[0013] The classifier is configured to generate predicted target parameters for the generator based on the scattering data generated by the generator.

[0014] The training process of the generative adversarial network includes:

[0015] Training the discriminator and classifier: Fix the network parameters of the generator, use the training sample data as real samples, use the scattering data generated by the generator as fake samples, combine the real samples and fake samples according to a predetermined ratio, and input them into the discriminator and classifier for iterative training until the training ends.

[0016] Training the generator: Fix the network parameters of the discriminator and classifier, and use the scattering data generated by the generator as real samples to input into the discriminator and classifier for iterative training until the training termination condition is met.

[0017] Furthermore, when training the discriminator and classifier, the ratio of real samples to fake samples is 1:1.

[0018] Furthermore, when training the discriminator and classifier, the training sample data is obtained by modeling and simulating vortex electromagnetic waves incident on a target with given target parameters.

[0019] Furthermore, during the training process of the generative adversarial network, stochastic gradient descent is used to update the network parameters.

[0020] Furthermore, the discriminator and classifier constitute an autoencoder network.

[0021] Furthermore, the discriminator and classifier share a sequentially connected input layer, at least one convolutional layer, a flattening layer, and a first fully connected layer, which is followed by a Dropout layer; each convolutional layer and the first fully connected layer are followed by a LeakyReLU activation layer; the discriminator is also connected to a second fully connected layer with an output dimension of 1 after the first fully connected layer, which is followed by a Sigmoid activation layer; the classifier is also connected to a third and a fourth fully connected layer after the first fully connected layer, which reduce the output dimension to be the same as the target parameter.

[0022] Furthermore, the generator includes an input layer, a connection layer, a fifth fully connected layer, a batch normalization layer, a sixth fully connected layer, a reshaping layer, and at least one convolutional layer connected in sequence; the fifth fully connected layer is followed by a LeakyReLU activation layer.

[0023] Furthermore, real and fake samples are distinguished by binarizing the labels on the sample data.

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0025] 1. Existing technologies mainly address the problem of predicting the electromagnetic scattering characteristics of targets under plane wave illumination. This invention uses vortex electromagnetic wave scattering data to train a network model, filling the gap in the prediction of vortex electromagnetic wave target scattering characteristics and providing a foundation for the widespread application of vortex electromagnetic waves in radar detection, identification and other fields.

[0026] 2. This invention omits the iterative correction process of the target physical model in the traditional modeling process that relies on high-frequency modeling and simulation technology. It can quickly and effectively map the target's geometric parameters to electromagnetic scattering characteristic data, and can complete the acquisition of scattering data in a shorter time, thereby improving the efficiency of data acquisition and reducing costs.

[0027] 3. This invention requires only a small memory overhead to achieve real-time acquisition of target electromagnetic scattering characteristic data within the range of geometric parameter variations, with minimal resource consumption. Attached Figure Description

[0028] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0029] Figure 1 This is a schematic diagram of the network architecture of the generative adversarial network model used in this invention.

[0030] Figure 2 This is an example of a hierarchical structure design for a generative adversarial network model.

[0031] Figure 3 This is a schematic diagram of the training process for a generative adversarial network model.

[0032] Figure 4 It is a set model diagram of the test targets.

[0033] Figure 5 This is a comparison chart of the predicted scattering data and the simulated data for targets with different parameters. Detailed Implementation

[0034] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0035] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0036] This invention provides a method for predicting the scattering characteristics of vortex electromagnetic waves based on generative adversarial networks. This method mainly addresses the following issues:

[0037] (1) Most existing methods for predicting electromagnetic scattering characteristics are mainly aimed at the electromagnetic scattering characteristics of targets irradiated by plane waves, and are not applicable to the prediction of scattering characteristics irradiated by vortex electromagnetic waves. This invention mainly uses generative adversarial networks to predict the scattering characteristics of vortex waves on metal targets. Through training, the network model can predict the electromagnetic scattering results of the same type of metal targets under vortex wave irradiation with different sizes and parameters.

[0038] (2) Current AI-based electromagnetic scattering prediction methods typically require a large amount of data samples and computational resources to enable the network model to learn the patterns in the electromagnetic scattering process. To address the issues of large data requirements and the time-consuming process of obtaining data through simulation, this invention directly generates electromagnetic scattering data of the same type of metal target under vortex wave irradiation with different sizes and parameters through a generative adversarial network model, thus meeting the data requirements for network training in vortex wave target identification and detection.

[0039] High-frequency modeling and simulation techniques are used to generate scattering data of vortex electromagnetic waves at a target. This data is then used to train a generative adversarial network (GAN), enabling the prediction of the electromagnetic scattering characteristics of vortex electromagnetic waves at targets with different parameters based on the input target parameters. The simulated target has specific parameters (such as geometric dimensions), and the simulated scattering data corresponds accordingly to these parameters.

[0040] Prediction methods include:

[0041] The target parameters and random noise are input into a trained generative adversarial network to generate vortex electromagnetic wave scattering data of the target. The so-called target parameters are those that have a decisive influence on the scattering characteristics, such as target geometric dimensions, vortex wave topological charge, and semi-cone angle parameters.

[0042] like Figure 1 As shown, a generative adversarial network (GAN) consists of a generator G, a discriminator D, and a classifier Q. Wherein:

[0043] The generator G takes two inputs: random noise z and target parameters c. Based on the input target parameters c and random noise z, the generator G generates vortex electromagnetic wave scattering data about the target.

[0044] The discriminator is configured to distinguish between true and false input training sample data or scattering data generated by the generator. Therefore, the output of the discriminator is a binary value, which outputs 1 when the judgment is true and 0 when the judgment is false.

[0045] The classifier is configured to generate predicted target parameters for the generator based on the scattering data generated by the generator, for use by the generator when it generates scattering data again.

[0046] Therefore, similar to traditional Generative Adversarial Networks (GANs), the generator G aims to generate scattering data similar to the real training sample data x, ensuring that the generated scattering data possesses similar statistical characteristics to the training sample data. The discriminator D, on the other hand, strives to distinguish the generated scattering data from the real training sample data. Furthermore, guided by the target parameter c, the GAN in this embodiment achieves fine-grained control over the generated scattering data, enabling the generator G to generate scattering data with specified attributes (corresponding to a specified target scene).

[0047] In some embodiments, the following are defined:

[0048] The input to generator G is (batch_size, noise_dim + discrete_dim + continuous_dim), where batch_size represents the batch size, i.e., the number of samples trained in each iteration; noise_dim is the dimension of the input random noise; discrete_dim is the dimension of the discrete target parameters; and continuous_dim is the dimension of the continuous target parameters. The output of generator G is (batch_size, channel, img_cols, img_rows), where channel represents the number of channels in the scattering data; and img_cols and img_rows represent the column and row dimensions of the generated scattering data, respectively.

[0049] The input to discriminator D is (batch_size, channel, img_cols, img_rows), and the output of discriminator D is (batch_size, 1).

[0050] The input to classifier Q is (batch_size, channel, img_cols, img_rows), and the output of classifier Q is (batch_size, discrete_dim + continuous_dim).

[0051] like Figure 2 As shown, in some embodiments, the hierarchical structure of the generative adversarial network is designed as follows:

[0052] Since the inputs to discriminator D and classifier Q are both (real or fake) scattering data, discriminator D and classifier Q are constructed as an autoencoder network, with discriminator D acting as the encoder and classifier Q acting as the decoder.

[0053] like Figure 2 As shown in (a), the discriminator D and the classifier Q share a sequentially connected input layer, at least one convolutional layer, a flattening layer, and a first fully connected layer (Dense), which is followed by a Dropout layer. Each convolutional layer and the first fully connected layer are followed by a LeakyReLU activation layer. These are the shared layers. The unique features of the two are as follows: the discriminator D is followed by a second fully connected layer with an output dimension of 1 after the first fully connected layer, which is followed by a Sigmoid activation layer; the classifier Q is followed by a third and a fourth fully connected layer after the first fully connected layer, which reduce the output dimension to the same as the target parameter.

[0054] like Figure 2 As shown in (b), the generator G comprises an input layer, a concatenate layer, a fifth fully connected layer, a batch normalization layer, a sixth fully connected layer, a reshape layer, and at least one convolutional layer (Conv) connected in sequence; a LeakyReLU activation layer is connected after the fifth fully connected layer. The introduction of the batch normalization layer plays an important role in the training of the generator G, as it helps to keep the gradient within a certain range, thereby reducing the possibility of gradient vanishing or gradient exploding, and improving the stability of training during backpropagation. This design can accelerate the convergence speed of the generator G, while improving the generator G's ability to learn the distribution of input data, thus producing more realistic and diverse samples.

[0055] exist Figure 2 In the illustrated embodiment, both the generator G and the discriminator D (shared with the classifier Q) contain only one convolutional layer.

[0056] like Figure 3 As shown, the training process of a generative adversarial network includes:

[0057] The discriminator D and classifier Q are trained. Discriminator D and classifier Q can be collectively referred to as the discriminant network: The network parameters of generator G are fixed. The training sample data x is used as the real sample and the scattering data generated by generator G is used as the fake sample. The real sample and the fake sample are combined according to a predetermined ratio and then input into discriminator D and classifier Q for iterative training until the training termination condition is reached.

[0058] Training the generator G: Fix the network parameters of the discriminator D and the classifier Q, and use the scattering data generated by the generator G as real samples to input into the discriminator D and the classifier Q for iterative training until the training termination condition is met.

[0059] The training sample data x for training the discriminator D and classifier Q is obtained through high-frequency modeling simulation. This simulation involves modeling a vortex electromagnetic wave incident on a target with given parameters to obtain the training sample data x. Unlike traditional prediction methods, this high-frequency modeling simulation is only used to obtain the training sample data x, rather than relying entirely on it for prediction. Furthermore, the ratio of the training sample data x to the scattering data generated by the generating gas G is typically 1:1 when input into the discriminator network.

[0060] For training the discriminator network, it is necessary to inform the discriminator D whether the input is a real sample or a fake sample. This can be achieved by labeling the input data. Similar to how the discriminator D represents the discrimination result, real and fake samples can be represented by binarizing the labels on the sample data, with a label of 1 representing a real sample and a label of 0 representing a fake sample.

[0061] Whether training the discriminator network or the generator G, the process involves adjusting the network parameters of the generative adversarial network by comparing the performance of the generated scattering data with that of the training sample data x in the discriminator network. Methods for adjusting the network parameters can include stochastic gradient descent. Through continuous iteration, the network training is complete when an equilibrium state is reached between the discriminator network and the generator G (i.e., network parameter convergence). The trained generative adversarial network then possesses the ability to predict the scattering characteristics of targets with different target parameters when vortex electromagnetic waves are used as the incident source. Furthermore, it can quickly generate electromagnetic scattering data of targets under different target parameters using vortex electromagnetic waves.

[0062] like Figure 4 The diagram shows the structure of a test target, a simple model composed of a large rectangle (length a, width b, height h), a small cuboid (length a / 2, width b, height h), and a cylinder (height a / 2, diameter R) arranged from bottom to top. First, scattering datasets with different target parameters were generated using joint simulations with Matlab and CST. By modifying the simulation model parameters, 1296 training samples (360*1 pixels) x were obtained to train the generative adversarial network. After the network training is complete, different scattering feature prediction results can be obtained by inputting different target parameters. Figure 5 The comparison results of simulated data and network prediction data at different altitudes (h) are presented. The Kolmogorov-Smirnov (KS) test (as shown in Table 1) indicates that the predicted data and simulated data belong to the same distribution, verifying the effectiveness of the prediction model.

[0063] Table 1. KS Distance between Predicted and Simulated Data at Different h Levels

[0064]

[0065] Furthermore, this embodiment also compares the efficiency of the prediction method of the present invention with that of traditional high-frequency modeling and simulation prediction. Table 2 shows the efficiency of the prediction data generated by the present invention (deep learning) and high-frequency modeling and simulation (high-frequency simulation). As can be seen from Table 2, the present invention achieves a significant improvement in prediction efficiency when calculating the same amount of data.

[0066] Table 2 Comparison of data generation efficiency between high-frequency modeling simulation and the present invention.

[0067]

[0068] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for predicting scattering characteristics of a target of a whirling electromagnetic wave based on a generative adversarial network, characterized in that, The application relates to a method for generating scattering data of a target under vortex electromagnetic wave. The method comprises the following steps: inputting target parameters and random noise into a trained generative adversarial network to generate scattering data of the target under the vortex electromagnetic wave; wherein the generative adversarial network comprises a generator, a discriminator and a classifier; the generator is configured to generate the scattering data of the target under the vortex electromagnetic wave according to the inputted target parameters and random noise; the discriminator is configured to distinguish the authenticity of the inputted training sample data or the scattering data generated by the generator; the classifier is configured to generate predicted target parameters to the generator according to the scattering data generated by the generator; the training process of the generative adversarial network comprises the following steps: training the discriminator and the classifier: fixing the network parameters of the generator, taking the training sample data as real samples and the scattering data generated by the generator as false samples, combining the real samples and the false samples according to a predetermined proportion, and inputting the combined samples into the discriminator and the classifier for iterative training until a training end condition is reached; 2. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, training the generator: fixing the network parameters of the discriminator and the classifier, taking the scattering data generated by the generator as real samples, and inputting the real samples into the discriminator and the classifier for iterative training until a training end condition is reached.

3. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, When training the discriminator and the classifier, the proportion of the real samples to the false samples is 1:

1.

4. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, When training the discriminator and the classifier, the training sample data is obtained by modeling and simulating the vortex electromagnetic wave incident to a target with given target parameters.

5. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, In the training process of the generative adversarial network, the random gradient descent method is used to update the network parameters.

6. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 5, characterized in that, The discriminator and the classifier constitute a self-encoding network.

7. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, The discriminator and the classifier share an input layer, at least one convolutional layer, a flattening layer and a first full connection layer connected in sequence, and the first full connection layer is connected with a Dropout layer; each convolutional layer and the first full connection layer are connected with a LeakyReLU activation layer; the discriminator is further connected with a second full connection layer with an output dimension of 1 after the first full connection layer, and the second full connection layer is connected with a Sigmoid activation layer; the classifier is further connected with a third full connection layer and a fourth full connection layer after the first full connection layer, and the third full connection layer and the fourth full connection layer reduce the output dimension to the same as the target parameters.

8. The vortex electromagnetic wave target scattering characteristic prediction method based on a generative adversarial network according to claim 1, characterized in that, The generator comprises an input layer, a connection layer, a fifth full connection layer, a batch normalization layer, a sixth full connection layer, a reshaping layer and at least one convolutional layer connected in sequence; and the fifth full connection layer is connected with a LeakyReLU activation layer. The real samples and the false samples are realized by binary labeling the labels of the sample data.

Images